The U.S. Army Cyber Command is working on a system that would allow AI agents to independently infiltrate enemy computer networks, seek out targets, and disable them. The military wants the AI to perform such tasks with little human intervention. However, experts fear that the agents could also attack systems that were not their targets.
The U.S. Army Cyber Command is working on a system that would allow AI agents to independently infiltrate enemy computer networks, seek out targets, and disable them. The military wants the AI to perform such tasks with little human intervention. However, experts fear that the agents could also attack systems that were not their targets.
This was reported by The Intercept, citing documents that the US military provided to potential contractors.
The project is codenamed ECHOTRIBBLE. According to the developers, the system will work like a computer worm — a malicious program that spreads between devices. For example, if an AI agent gets on a laptop, it can find other devices on the same network and create its own copies, adapted to work on them. The military presentation gives an example of an agent self-propagating to an iPad, an Apple computer, and a Windows 11 PC.
Agents should also communicate with the military as rarely as possible, hide data transmissions, and delete themselves after completing the task to leave no traces. In addition, the system should work in cloud services and even isolated systems that are not connected to the Internet.
Humans will be able to set the conditions for the attack and identify targets in advance. However, the documents explicitly state that the system must operate with minimal human control and make decisions independently.
However, experts warn that an AI agent can achieve its goal, but at the same time cause damage to third-party systems. Hyrum Anderson, CEO of cybersecurity company Abundant, gives an example: what if the agent sees that the enemy is managing its operations through an American water utility? Wouldn't the AI decide to disable it too?
Previously, AI agents have tried to infiltrate computer systems during tests, even though no one instructed them to do so. According to researchers, if such systems are directly ordered to hack and destroy digital infrastructure.
Previously, dev.ua wrote that OpenAI's AI agents were trying to obtain information from US government websites without permission.
